Celebrity Hairstyle Recommendation System with Hairstyle Transfer

Hairstyles significantly shape an individual’s personality and fashion sense, accentuating appearance in social perception. The challenge of selecting a suitable hairstyle for one’s face is crucial as mass media expands access to celebrity fashion trends. Hairstyle recommendation systems have been introduced; however, they exhibit certain drawbacks, such as relying on facial shape classification combined with rule-based methods to match specific hairstyles to user face shapes. This paper introduces a Celebrity Hairstyle Recommendation System with Hairstyle Transfer, which will provide more flexible and personalized hairstyle options for the user’s face. Our method first uses the Swinface model to extract and analyze the facial features. Secondly, we apply a comparison algorithm based on the similarity between the attributes of the user’s face and the faces of celebrities in the database to recommend the four most suitable hairstyles. Finally, the new hairstyles will be quickly and accurately rendered on the user’s face using the HairFastGAN model. We conduct experiments by applying our recommendation algorithms on the SCUT-FPB5500 dataset, using a SEResNeXt50 backbone combined with ComboLoss to score the input and generated images to determine an effective recommendation strategy. The experimental results indicate that our recommendation method improves the average beauty score by 7.69% for the faces in the filtered SCUT-FBP5500 dataset, with the best PC metric of approximately 0.96.